Journal of Experimental Child Psychology 79, 139–161 (2001)

doi:10.1006/jecp.2000.2574, available online at http://www.idealibrary.com on

Modeling Face Identification Processing

in Children and Adults

Gudrun Schwarzer

University of Tübingen, Tübingen, Germany

and

Dominic W. Massaro

University of California, Santa Cruz

Two face identification experiments were carried out to study whether and how children
(5-year-olds) and adults integrate single facial features to identify faces. Using the paradigm
of the Fuzzy Logical Model of Perception each experiment used the same expanded factori-
al design, with three levels of eyes variations crossed with three levels of mouth variations as
well as their corresponding half-face conditions. In Experiment 1, an integration of facial fea-
tures was observed in adults only. But, in adjusting the salience of the features varied, the
results of Experiment 2 indicate that children and adults evaluated and integrated informa-
tion from both features to identify a face. A weighted Fuzzy Logical Model of Perception fit
the judgments significantly better than a Single Channel Model and questions previous
claims of holistic face processing. Although no developmental differences in the stage of the
integration of facial information were observable, differences between children and adults
appeared in the information used for face identification. © 2001 Academic Press
Key Words: face perception; information processing; perceptual development.

Notwithstanding the impressive face identification capabilities during infancy

(e.g., Bushnell, 1982; Bushnell, Sai, & Mullin, 1989; Pascalis, de Haan, Nelson,
& de Schoenen, 1998), face identification during the 1st decade of life continues
to undergo development. Young children are dramatically worse than adults at
encoding and subsequently identifying unfamiliar faces. Marked improvement
between ages 2 and 10 is observed on simple recognition tasks (for an overview
see Flin & Dziurawiec, 1989). Although these differences could be differences in
information processing of faces, current research does not offer a definite answer.
On the one hand, the literature on face recognition suggests that 6-year-old chil-
dren as well as adults process faces holistically (e.g., Carey, 1996). On the other

Address correspondence and reprint requests to Gudrun Schwarzer, Friedrich-Miescher-Laboratory

of the Max-Planck-Society, Spemannstr. 34, 72076 Tübingen, Germany. E-mail: gudrun.schwarzer@
tuebingen.mpg.de.

139
0022-0965/01 $35.00
Copyright © 2001 by Academic Press
All rights of reproduction in any form reserved.
140 SCHWARZER AND MASSARO

hand research on nonfacial visual processing modes demonstrated that children 5

years of age and older process visual stimuli analytically as do adults (Ashkenasy
& Odom, 1982; Thompson & Massaro, 1989; Wilkening & Lange, 1989).
In an attempt to integrate these research lines, in the present study, we investi-
gate face processing by employing expanded factorial designs varying several
sources of information and mathematical model testing. We examine whether
children’s and adults’ face processing can be explained by analytic models of per-
ception like the Fuzzy Logical Model of Perception or the Single Channel Model.
If these models are able to explain face identification data, evidence for analytic
face processing is given; if they fail to explain the data, analytic face processing
would be questionable.

Research on Processing of Faces in Adults

Most recent research on face identity processing focused on issues of holistic
and analytic processing (Bruce & Humphreys, 1994; Diamond & Carey, 1986;
Farah, Tanaka, & Drain, 1995; Farah, Wilson, Drain, & Tanaka, 1998; Tanaka &
Farah, 1993; Tanaka & Sengco, 1997). However, whereas most investigators
agree on the definition of analytic processing in terms of analyzing featural
dimensions, at least two definitions of holistic processing can be differentiated:
holistic encoding and configurational encoding (Carey & Diamond, 1994).
According to Tanaka and Farah (1993) as well as Farah et al. (1998), holistic
encoding means that a particular face is not represented in terms of its parts but
in terms of a templatelike representation of the whole face. Farah, Tanaka, and
Drain (1995) characterized holistic representations as relatively non-part-based
representations. In contrast, configurational encoding emphasized the specific
spatial relationships among the individual facial features (first-order relational
information) as well as the second-order relationships which include the spatial
locations of the facial parts relative to the prototypical arrangement of the eyes,
nose, and mouth (Diamond & Carey, 1986). Diamond and Carey suggested that
with experience, participants develop a finely tuned face prototype, with respect
to which they can encode the second-order relational information in faces.
Both kinds of holistic processing are in agreement with the findings of Tanaka
and Farah (1993) that individual facial features were more easily recognized
when they were part of the complete face than when they were presented alone.
In Tanaka and Farah’s task, participants learned to identify a set of faces and a set
of some contrasting class of stimuli (scrambled faces, inverted faces, and hous-
es). The task was to identify the single parts of the faces in isolation as well as in
the context of the whole face. The results showed that participants identified
upright facial stimuli more easily and correctly in the whole face condition,
whereas this advantage disappeared with other classes of visual stimuli as scram-
bled faces, inverted faces, or houses.
A current version of holistic processing also acknowledges the influence of
individual facial parts (Tanaka & Sengco, 1997). The authors demonstrated that
adults performed above chance when recognizing face parts presented in isolation,
 FACE PROCESSING 141

indicating that the individual face parts were encoded independently of the other
features and their configuration. Nevertheless, the authors emphasized the inter-
action between featural and configurational information from a face because their
results also showed that alterations in facial configurations interfered with the
retrieval of facial features, whereas this interference did not appear with inverted
faces or nonface stimuli. Thus, for adults, current research suggests that upright
faces are processed more holistically than inverted faces or other nonface stimuli.

Research on Face Processing in Children

Tanaka, Kay, Grinell, Stansfield, and Szechter (1998) argued for holistic proc-
essing because they observed an advantage of identifying facial parts in whole
faces relative to those presented in isolation during the age range between 6 and
10 years. In line with these results, the studies carried out by Baenninger (1994)
emphasized young children’s (6 years of age) tendency to rely more on configur-
al information than on featural information. In light of new findings, Carey and
Diamond (1994) and Carey (1996) modified their original assumption of an ana-
lytic to holistic shift. Using the task of Young, Hellawell, and Hay (1987), Carey
and Diamond (1994) asked children between 6–7 and 10 years as well as adults
to name the upper half of a face when the lower half of the face belonged to a dif-
ferent person. In different conditions the face halves were aligned (composite
faces) or not aligned (noncomposite faces) and the faces were shown upright or
inverted. The authors reported that children and adults were slowed by the mis-
match of the upper half of the composite in comparison to the noncomposite faces
and at all ages this interference was obtained only when the faces were upright.
Thus, there was no processing shift from analytic to holistic processing because
at all ages children processed the faces holistically. The effect of inversion on the
measure of holistic processing was constant over age. Second, older participants
were more affected by inversion than were younger participants. With increasing
age inverted faces were encoded less efficiently than upright faces. This latter
effect was independent of the effect of inversion on the measure of holistic encod-
ing. In sum, the current state of affairs concerning the development of face pro-
cessing is that children as young as 6 years of age process upright faces but not
inverted faces holistically, just as do 10-year-olds and adults.

Research on Processing of Nonfacial Stimuli in Children and Adults

The conclusion that children and adults process faces holistically stands in
sharp contrast to findings on visual development using nonfacial stimuli. A long-
standing proposal is centered around the idea of a developmental shift from holis-
tic modes of processing in young children (around age 4 to 7) to analytic modes
of processing in older children and adults (Kemler Nelson, 1989; Shepp, 1978;
Smith & Kemler, 1977; Werner, 1926). Although this view of a developmental
shift has been challenged and modified, subsequent studies did not question the
assumption of an adult as an analytic processor but simply the proposal that chil-
dren are typically holistic processors.
142 SCHWARZER AND MASSARO

One issue concerns the diagnosis of holistic processing in the commonly used
restricted classification task (Ashkenasy & Odom, 1982). The authors proved
that the putative holistic classifications could be the result of not finding an exact
match on the dimension young children focused on. Indeed, Ashkenasy and
Odom (1982) found that young children focused on single visual dimensions,
usually the dimension containing the most stimulus variability or distinctiveness
(see also Odom & Guzman, 1972). In a differential sensitivity account of cogni-
tive development, Cook and Odom (1992) explicated that what changes across
development is not the processing mode—from holistic to analytic—but the
information’s relative position in hierarchies of perceptual sensitivity.
Performance is more accurate in problems that require use of dimensions that are
higher in perceptual salience than in problems that require use of dimensions
lower in salience. Salience can change with prior experience with the dimension
that is perceived and the way the dimension is physically represented in the envi-
ronment.
An important concern is that the diagnosis of holistic processing in children
has analyzed average results and ignored individual differences. When individual
results are analyzed in the context of the restricted classification task paradigm or
a concept-learning task, children and also adults have a strong bias to use just a
single dimension to make their judgments (e.g., Schwarzer, 1997; Thompson,
1994; Ward, Vela, & Hass, 1990; Wilkening & Lange, 1989).
Another line of research explored the nature of perception in preschool chil-
dren and adults during the earliest moments of visual processing (Thompson &
Massaro, 1989). In contrast to the mentioned restricted classification or concept
learning tasks, the goal was to investigate perceptual processing while minimiz-
ing processes of postperceptual decision making. The experimental test was car-
ried out within the paradigm of a pattern recognition situation and mathematical
model testing within the context of the Fuzzy Logical Model of Perception
(explained in detail below). The general logic of these experiments is to manipu-
late the features of a multidimensional stimulus independently of one another and
to measure the rule of their joint influence on the performance of interest. In the
study by Thompson and Massaro (1989) the children’s and adults’ judgments
were best described by the predictions of the Fuzzy Logical Model of Perception.
The model predicts that children and adults encode (evaluate) single features
independently and combine them during a second multiplicative operation.
Independent encoding of features is consistent with analytic processing rather
than holistic processing. The role of individual features in the conceptualization
of holistic processing—if they have any role—is only secondary and processing
of the whole stimulus cannot be reduced to processing of separate features (see
Kemler Nelson, 1989).
In sum, previous studies on visual nonfacial processing concerning the putative
holistic processing in children showed that children, like adults, indeed have
access to single visual features and therefore their processing modes can be
named as analytical.
 FACE PROCESSING 143

There is evidence that children process nonfacial visual stimuli analytically

and also that the processing of faces is built up from single facial attributes
(Schwarzer, 2000; Schwarzer & Korell, in press). These authors used a category
learning task in which children and adults had to categorize faces which varied in
four attributes (eyes, nose, mouth, and outline). The construction of the categories
allowed participants to categorize faces either by focusing on a single facial
attribute or by taking the features of the whole face into account. The results
showed that children of ages 4–5 and 10 years categorized faces by focusing on
single facial features, whereas the majority of the adults based their categoriza-
tions on the whole face, using three or four facial attributes. Moreover, Schwarzer
and Korell’s (in press) study is the first to show that even 4- to 5-year-old chil-
dren are able to recognize faces above a chance level. By using an age-specific
experimental setting Schwarzer and Korell (in press) demonstrated significant
face recognition capabilities even in processing very similar faces. Moreover,
Schwarzer’s results (2000) were consistent with those from Pedelty, Levine, and
Shevell’s (1985) study on face processing using a multidimensional scaling
approach. Young children (between 7 and 10 years of age) focused on singular
dimensions, whereas older children and adults attended to two or more dimen-
sions when making face similarity judgments. Thus, developmental differences
seem to appear only with respect to the amount of facial information (one feature
or more features) used in processing and not with respect to the mode of pro-
cessing (analytic or holistic). However, these studies do not address the question
of how multiple facial features are integrated when several features are used.
These studies as well as the studies on nonfacial visual processing cast doubt
on the conclusion that children and adults process faces holistically. One appro-
priate method to test the hypothesis of analytic and holistic processing—not used
in the context of face processing until now—is to test whether analytic models
like the Fuzzy Logical Model of Perception or the Single Channel Model could
explain face identification in children and adults. If these models fit data on face
identification, evidence for analytic processing is given; if they fail to explain the
data holistic face processing seems more appropriate.

Fuzzy Logical Model of Perception

The Fuzzy Logical Model of Perception assumes three stages between stimu-
lus event and a response: evaluation, integration, and decision. With regard to the
current experiments, the Fuzzy Logical Model of Perception predicts that partic-
ipants evaluate the identity of a face according to information arriving from mul-
tiple sources, i.e., separate facial features. Stimulus variation is made along sev-
eral dimensions [in this case, eyes variations (including variations of eyebrows
and forehead) and mouth variations]. An important criterion for manipulating two
features is that they can be varied independently of one another. Varying one cue
in the upper face and one cue in the lower face fulfilled this condition. In the cur-
rent experiments, three levels of the upper face and three levels of the lower face
were combined factorially. In addition, according to the expanded factorial
144 SCHWARZER AND MASSARO

design, six half-face conditions were presented. The half-face conditions dis-
played only the three levels of the upper and the three levels of the lower halves
of the stimulus faces. The variations of the eyes (including their covariates) and
mouth features were varied as can be seen in Fig. 1.1
The eyes, eyebrows and the height of the forehead were varied in the upper face.
As is shown in Fig. 1 (comparing the faces within each column), the eyes varied
from straight gaze, narrow eyebrows, and a long forehead (Level 3 at the bottom of
each column) to a gaze toward the left corner, wider eyebrows, and a shorter fore-
head (Level 1 at the top of each column). Variations of the mouth in terms of its
width comprised the second factor. The mouth varied from a small mouth (Level 1
at the left side of each row) to a wide mouth (Level 3 at the right side of each row).
In order to use the 15 faces (9 whole faces plus 6 face-halves) in the context of
a face identification task we defined two prototypical faces. These faces con-
tained the extreme levels on both features, eyes (and their covariates) and mouth.
One prototype had eyes with a straight gaze, a long forehead, narrow eyebrows,
and a wide mouth (prototype with Level 3 for the eyes and Level 3 for the mouth;
see the face at the right corner at the bottom). The contrasting prototype had eyes
which look to the left corner, a short forehead, wider eyebrows, and a small
mouth (prototype with Level 1 for the eyes and Level 1 for the mouth, see the face
at the left corner at the top). We named these prototypes Bob (1,1) and John (3,3).
The participants had been familiarized with these prototypes and the prototypes
were present during the entire experiment.
Within the context of the Fuzzy Logical Model of Perception, it is assumed that
participants evaluate the information from each source (i.e., the varied features)
according to the degree of match to the given prototypes. At the feature evalua-
tion stage, each physical input is transformed to a psychological value and is rep-
resented in the model equations in lowercase (e.g., if Ei represents the eyes infor-
mation, Ei would be transformed to ei, the degree to which the eyes variations
support the alternative “Bob”). An important assumption is that the evaluation of
a particular feature occurs independently of the presence or absence of other fea-
tures and their information value. With just two alternatives, Bob (B) and John (J),
we can make the assumption that the degree to which the eyes variations (includ-
ing their covariates) support alternative J is 1 − ei (Massaro & Friedman, 1990).
Feature evaluation occurs analogously for the mouth variations, Mj.
The evaluation stage is followed by the integration stage. Feature integration
consists of a multiplicative combination of the feature values supporting a given
alternative. If ei and mj are the values supporting the alternative B, then the total
support, M(B), for the alternative B would be given by the product of ei and mj:
M(B) = eimj.
The third stage within the Fuzzy Logical Model of Perception is decision,
which uses a relative goodness rule (Massaro & Friedman, 1990) to give the

1
We thank Dr. N. Troje for constructing and providing the faces.
 FACE PROCESSING 145

FIG. 1. Stimulus faces used in Experiment 1. The four faces displaying the maximum feature
variations (at the corners of the figure) as well as faces displaying “neutral” variations (in the mid-
dle of the figure). The prototype “Bob” is the face at the left corner at the top and the prototype
“John” is the face at the right corner at the bottom. The half-face conditions displayed only the
upper or lower half of the stimulus face. The reader can cover half of the face to experience these
conditions.

relative degree of support for each of the test alternatives. In the two-alterna-
tive choice task, the probability of a Bob choice, P(B), is equal to
P(BⱍEiMj) = M(B)/[M(B) + M(J)],
where P(BⱍEi,Mj) is the predicted choice given stimulus Ei, Mj. The Fuzzy
Logical Model of Perception requires three free parameters for the eyes variations
and three for the mouth variations. The parameters represent the degree to which
these features match those in the Bob and John prototypes.
146 SCHWARZER AND MASSARO

Single Channel Model

The question of analytic or holistic processing could also be answered if alter-
native mathematical models, especially nonintegrative models like the Single
Channel Model, fit the observed data. This is the case because the Single Channel
Model emphasizes the independent evaluation of single features, which is the
antithesis of holistic models, as is the Fuzzy Logical Model of Perception.
Whereas the Fuzzy Logical Model of Perception assumes an integration of the fea-
ture values, it is assumed in the Single Channel Model that only one of the inputs
of the upper or lower part of the face is functional on any trial. The information on
the upper part of the face is selected with some bias probability p and the lower
part with bias 1 + p. Using Bob as an example, for a given whole-face condition
the upper face information, Ei, will be identified as one specific face, with proba-
bility ei, and the lower face information with probability mj. Thus, the predicted
probability of the identification of Bob given the ith level of the upper face infor-
mation, Ei, and the jth level of the lower face information, Mj, is as follows:
P(BⱍEiMj) = (p)(ei) + (1 - p)mj.
This equation predicts the probability of identifying Bob for each of the nine
conditions in the factorial conditions. The predictions for the half-face conditions
are given simply by ei and mj. Since the 15 equations have three different values
of ei and three different values of mj and we also do not know the value of p in
the whole-face conditions seven free parameters are necessary: the p value, the
three ei values, and the three mj values.
The Single Channel Model is a nonintegration model that assumes that only
one of multiple inputs is used. In contrast, the Fuzzy Logical Model of Perception
is an integration model that proposes a multiplicative combination of several fea-
tures. Thus, examining the fit of the Fuzzy Logical Model of Perception in com-
parison to the fit of the Single Channel Model answers not only the general ques-
tion of analytic or holistic processing but also the specific question of whether
and how the facial features are integrated.

Models of Holistic and Configurational Encoding

Unfortunately, we know of no holistic model that can be quantitatively tested
against the results. Holistic models in terms of holistic encoding would assume
that each unique feature combination would create a unique face that could not
be predicted from its component features. This formulation captures the idea of
Garner’s (1974) description of integral (holistic) perception in that features do not
have any psychological reality in integral perception. According to Garner, fea-
tures of an integral stimulus only exist in the experimenter’s mind and not in the
mind of the perceiver. It is not possible to test a specific quantitative formulation
of this holistic model because it requires as many free parameters as observed
data points. Every face is unique and its identification cannot be predicted on the
basis of its components. This model remains untestable until there is some imple-
mentation of its principles with fewer free parameters. However, regardless of
 FACE PROCESSING 147

whether a particular holistic encoding model can be tested, an adequate fit of

either the Single Channel Model or the Fuzzy Logical Model of Perception pro-
vides evidence against holistic models because this would be evidence that face
identification is based on independently evaluated features.
The second characterization of holistic processing, configurational encoding,
refers to the possibility that the spatial relations among the parts of the face are
more influential than the parts themselves. The parts are represented but it is the
relation among the parts that are critical. In the framework of the Fuzzy Logical
Model of Perception, however, a relation between two parts of the face could
function as an additional source of information. Configurational encoding mod-
els would be most directly tested by manipulating the relations among parts of the
face. A good fit of the Fuzzy Logical Model of Perception would eliminate the
possibility that the spatial relation between the upper and lower half of the face
was a functional feature for face identification in our task. If spatial relations were
functional, the identification of the whole and half faces could not have been pre-
dicted with the same parameter values, which are necessarily based on the com-
ponent features. Tanaka and Sengco’s (1997) less strong version of holistic pro-
cessing based on the interaction of featural and configurational information also
predicts that the Fuzzy Logical Model of Perception should fail.
In summary, the expanded factorial design provides a strong test of the various
forms of holistic processing. The research paradigm is rich enough to investigate
what information is actually being used by the perceiver, even if a spatial relation
among features is involved.
In our experiments we test children’s (5-year-olds) and adults’ performance in
a face identification task against the predictions of the Fuzzy Logical Model of
Perception and the Single Channel Model. Although most of the face processing
studies focused on children as young as 6 years of age (e.g., Baenninger, 1994;
Carey, 1996), our study concentrated on the age group of 5-year-old children
because their face recognition capabilities are similar to the capabilities of 6-year-
old children (see Schwarzer & Korell, in press) and these results would be direct-
ly comparable to the studies from the nonfacial visual domain focusing on chil-
dren as young as 5 years of age (e.g., Ashkenasy & Odom, 1982; Thompson &
Massaro, 1989; Wilkening & Lange, 1989).
Taking into account that children’s perceptual sensitivity to the dimensions at
hand might be a significant factor for performance and development (see Cook &
Odom, 1992), we varied the distinctiveness of the attributes of the faces in two
separate experiments (1 and 2).

EXPERIMENT 1
Method
Participants. Two age groups participated in the experiment. One group was
composed of 10 5-year-old children. Six of them were girls, and four were boys.
Their ages ranged from 5 years, 0 months to 5 years, 6 months (mean age = 5,2)
and the children came from middle to upper middle-class families. Twenty-three
148 SCHWARZER AND MASSARO

students (mean age: 20 years; age range: 18–24) from the undergraduate psy-
chology subject pool at the University of California, Santa Cruz, made up the
adult group. All adults had normal, or corrected-to-normal, visual acuity. The
children were free of any known visual impairments.
Stimuli and apparatus. The nine stimulus faces of the experiment depicted in
Fig. 1 were constructed at the Max-Planck-Institut for Biological Cybernetics in
Tübingen, Germany, using the method of the correspondence-based representa-
tion of faces developed by Vetter and Troje (1997), which allowed for the con-
struction of continua along facial features. The four faces at the corners of Fig. 1
display the maximum variations of the features eyes (and their covariates) and
mouth. The faces in the middle row and column display neutral variations of the
features. The prototype “Bob” is the face at the left corner at the top and the pro-
totype “John” is the face at the right corner at the bottom. In addition to the nine
whole faces we constructed six face halves displaying only the three levels of the
upper and the three levels of the lower halves of the stimulus faces.
The experimental control programs used to run the experiment and collect par-
ticipant data (adults’ only) were implemented on a Silicon Graphics 4D/Crimson
Reality workstation running under the IRIX 5.3 operating system, the stimulus
faces were displayed to the participants on 12-in. (30.48 cm) NEC Model C12-
202A color monitors, and participants’ responses were collected on TV 950 video
display terminals (VDTs) and their associated keyboards. Data analyses were per-
formed on the same Silicon Graphics workstation using FORTRAN 77 data
analysis routines. To present this experiment at a preschool, the stimulus faces
were recorded on VHS tape so that the faces could be presented via a video
recorder. The data of the children’s group were collected manually.
Procedure. All participants were tested in a quiet location, with the experi-
menter present. They were seated in front of the computer monitors. Children
were tested in a research van parked outside of a day care center, and adults were
tested in a laboratory. Children and adults were required to respond to each stim-
ulus face either with Bob or John. Adults gave their response by pressing a cor-
respondingly labeled button on either the left or right edge of the VDT keyboard.
To minimize the effect of memory performance, pictures of the prototypes Bob
and John were fastened next to the response buttons. Children responded either
verbally (Bob or John) or by pointing to pictures of Bob and John that were fas-
tened beneath the monitor.
After a short title sequence, the control program began displaying the individ-
ual faces on the monitor. The faces were displayed for 1500 ms each. They were
sized to fill the vertical dimension of the 12-in. (30.48-cm) monitor screens (15
cm high) and were viewed at a distance of about 45 cm. No visual fixation point
was provided.
Each experimental session included a familiarization phase where the proto-
typical faces of Bob and John were presented 15 times each, 10 practice trials,
and 120 stimulus trials; the stimulus trials were selected from the stimulus set
according to a random selection without a replacement protocol, which resulted
 FACE PROCESSING 149

in each stimulus face being displayed 8 times per session (not including 15 famil-
iarization trials and 10 practice trials). Children and adults were involved in two
experimental sessions and so saw each stimulus face 16 times. In the adult group
the two experimental sessions were separated by a 5-min rest period during the
same day. For the children the two sessions were divided over two following
days. Each session per day was also divided into four subsessions where the chil-
dren had to identify 30 stimulus trials.

Results and Discussion

Mean probabilities of identifying “Bob.” Because in this experiment the par-
ticipants were limited to two choices, their mean responses to a particular stimu-
lus face could be expressed as a probability of identifying the faces as Bob [(P)B].
Consequently, the probability of identifying the faces as John was 1-(P)Bob. The
points in Figs. 2a and 2b show the average probabilities of identifying the faces
as Bob’s face as a function of the levels of the mouth and eyes variations in adults
and children respectively.
The left panels of Figs. 2a and 2b (points) show the identification of just the
lower half of the faces and the right panels for just the upper half. For the adult
group (see Fig. 2a), a comparable influence of eyes and mouth in the half-face
conditions was observable (see points of the left and right panels of Fig. 2a).
Also, in the whole-face conditions (see points of the middle panel of Fig. 2a) both
variables were significantly influential. There was also a significant interaction of
the variables in that the influence of one variable was larger than the extent to
which the other variable was neutral or ambiguous (all Fs were significant at the
.01 level).
In contrast, for the children (see Fig. 2b), the steeper course of the mouth vari-
able (see left panel) illustrates that the mouth was much more influential than the
eyes variable (see right panel), although both half-face conditions were effective
in changing the identification from Bob to John (Fs were significant at the .01
level). However, in the whole-face conditions (middle panel) the effect of the
mouth was significant only, F(2, 18) = 197.19, p < .01, whereas the influence of
the eyes did not reach the level of significance, F(2, 18) = 2.32, p > .05. This
means that the children’s identification of the whole face was nearly exclusively
influenced by the mouth. If the eyes were also influential, the points of the three
eyes levels would have been more separated from each other in the middle panel
of Fig. 2b.
Thus, a different use of facial information by adults and children was observed.
In the whole-face conditions, the adults took both sources of facial information
into account, whereas the children were solely influenced by the mouth of the
faces.
Model tests. The Fuzzy Logical Model of Perception makes specific predictions
about how participants should perform when viewing stimulus faces involving
features that are independent of one another. Identification probabilities should be
more extreme when features are congruent and nonambiguous, whereas they
150 SCHWARZER AND MASSARO

FIG. 2. (a) Predicted (lines) and observed [points; P(Bob)] identification judgments in the adults
in Experiment 1. Predictions are for the Fuzzy Logical Model of Perception. (b) Observed [points;
P(Bob)] identification judgments in the children in Experiment 1 as a function of the eyes and mouth
variations.
 FACE PROCESSING 151

should be less extreme when features disagree or are neutral or ambiguous.

Furthermore, as the ambiguity of one feature increases, judgments should be
influenced more by the other, less ambiguous feature.
For the model tests to answer the question of an analytic or holistic processing
as well as integrative or nonintegrative processing, both sources of information
have to influence performance. If there is an influence of just one source of infor-
mation the Fuzzy Logical Model of Perception and Single Channel Model make
the same predictions. Because, as described above, both sources of information
varied on the face had been influential in the adult group only, the model tests
were exclusively run in this age group (lines in Fig. 2a correspond to the predic-
tions of the Fuzzy Logical Model of Perception).
Model fitting was accomplished using the STEPIT subroutine (Chandler,
1969), which finds the parameter values that maximizes a model’s predictions. In
our tests, the subroutine minimized the root-mean-square deviation (RMSD)
between an individual participant’s observed data points and a set of data points
predicted by the Fuzzy Logical Model of Perception with six free parameters
(seven in the case of the Single Channel Model). The subroutine iterates the
model equations, changing parameter values until values are found that minimize
the RMSD between the observed data and the predicted data. The value of the
final RMSD is an indication of the model’s goodness of fit. The models are test-
ed against the results of individual participants in order to avoid the pitfalls in
analyzing average results and simultaneously allowing an exploration of individ-
ual differences (Massaro, 1998).
Given the 3 X 3 expanded factorial design, six free parameters are necessary to
fit the Fuzzy Logical Model of Perception to the 15 data points: three parameters
for each level of eyes and mouth. The parameters represent the degree to which
these features match those in the Bob prototype. As mentioned earlier, both the
fit of the Fuzzy Logical Model of Perception and the fit of the Single Channel
Model stand for an analytic processing in terms of an independent evaluation of
the facial features. If the Fuzzy Logical Model of Perception gives a better fit of
the data than the Single Channel Model, it can be argued that the facial features
were integrated by a multiplicative algorithm. Therefore, the Fuzzy Logical
Model of Perception and the Single Channel Model were fit to each of the adults
and to the mean participant computed by averaging the results across participants.
As can be seen in Fig. 2a (see lines), the predictions of the Fuzzy Logical Model
of Perception did reasonably well in capturing the trends in the data. The Fuzzy
Logical Model of Perception’s RMSDs for the adults ranged from .02 to .17, with
an average RMSD of .08. The fit of the mean participant gave an RMSD of .06.
The fit of the Single Channel Model produced larger RMSDs, with a range
between .02 to .22 and an average RMSD of .11. The fit of the Single Channel
Model to the mean participant data was .07.
A comparison of the RMSD fits of the two models is informative. The RMSD
fits for each subject for the two models were submitted to an analysis of variance,
and the Fuzzy Logical Model of Perception provided a significantly better overall
152 SCHWARZER AND MASSARO

fit than did the Single Channel Model, F(1, 22) = 4.86, p < .05. Thus, the good fit
of the Fuzzy Logical Model of Perception, which assumes independence at the
feature evaluation stage, provides evidence against holistic models. If each com-
bination of features is unique—as assumed by the holistic encoding approach—
then a model assuming independence between features should fail. The fact that
the Fuzzy Logical Model of Perception does not fail is therefore evidence against
holistic processing. Furthermore, the results of the present experiment showed
developmental differences in terms of the facial information used by children and
adults. Whereas the children nearly exclusively focused on the mouth of a whole
face, the adults took both the mouth and eyes into account. The children might
have been influenced exclusively by the mouth of the faces because they associat-
ed John’s mouth with a smiling mouth and managed the whole-face task by look-
ing for just the smiling face.
These results on the differences between children and adults in terms of the
facial information used fit nicely into Cunningham and Odom’s (1986) study on
differential salience of facial features in children’s perception of affective expres-
sion. Their results showed that 5- to 11-year-old children were more likely to
evaluate and remember information from the mouth region first and the eye
region second.
Experiment 1 did not allow to test different mathematical models of children’s
face identification capabilities because of the age related differences in using
facial information. Therefore, in Experiment 2 we changed the salience pattern of
the face stimuli to allow for model tests in children.

EXPERIMENT 2
To accomplish the goal to compare children’s and adults’ information process-
ing of faces by means of mathematical model testing, in Experiment 2, we mod-
ified some of the experimental conditions of Experiment 1 to get the children to
consider both facial features, eyes and mouth. We adjusted the salience of eyes
and mouth information given in the faces to increase the likelihood that the eyes
will be evaluated in the children as well. Salience is considered a characteristic of
the relative distinctiveness of the properties in stimulus arrays (Cunningham &
Odom, 1978). In Experiment 2, we decreased the salience of the mouth—the
most influential feature in the children’s responses observed in Experiment 1.
According to Ashkenasy and Odom (1982), we decreased the distinctiveness-
based salience of the mouth by decreasing the value differences of the mouth to
yield lower levels of sensitivity to these mouth differences. In addition, we
decreased the overall size of the stimulus faces to make sure that the whole face
could be projected on the fovea.

Method
Participants. There were 11 children from middle- to upper middle-class
families participants, 6 girls and 5 boys, ranging in age from 5 years, 4 months
to 5 years, 11 months (mean age: 5 years, 8 months). All children were attend-
 FACE PROCESSING 153

ing a kindergarten group at the time of testing. The 13 adults were students at
the University of California, Santa Cruz, 6 females and 7 males (mean age: 20
years; age range: 18–22). Participants had normal visual acuity, with or without
correction.
Stimuli. The stimuli were the same as in Experiment 1 with the following
exceptions. (1) To decrease the distinctiveness-based salience of the variable
mouth we increased the similarity of mouth Level 1 and 3 to Level 2 as can be
seen in Fig. 3. The variations of the eyes remained the same as in Experiment 1.
The stimulus faces of Experiment 2 consisted of nine new faces plus the corre-
sponding six face halves. We expected that in decreasing the distinctiveness-
based salience of the mouth we would increase the relative salience of the eyes
and therefore increase children’s sensitivity to the eyes variable. (2) To facilitate
the identification of Bob and John in the familiarization phase we constructed
new facial prototypes for Bob and John in which the features—mouth and eyes—
were exaggerated (see the faces of the separated right column of Fig. 3). The pro-
totypes were also shown on the pictures which were used by the participants to
give their responses during the test phase, but they were not used as test stimuli.
The test faces were the nine faces depicted in Fig. 3 plus the corresponding six

FIG. 3. Stimulus faces of Experiment 2. The four faces displaying the maximum feature variations
(at the corners of the figure) as well as faces displaying “neutral” variations. The prototype “Bob” is
the face at the top of the separated right column and the prototype “John” is the face at the bottom of
this right column. The half-face conditions displayed only the upper or lower half of the stimulus face.
154 SCHWARZER AND MASSARO

FIG. 4. Predicted (lines) and observed [points; P(Bob)] identification judgments in the children
(a) and adults (b) in Experiment 2. Predictions are for a weighted Fuzzy Logical Model of
Perception.
 FACE PROCESSING 155

face- halves. (3) We decreased the overall size of the faces (9 cm high) to encour-
age the participants’ consideration of both features varied. We expected that
smaller faces increase the probability that people look at the whole face.
Apparatus and procedure. These were the same as those used in Experiment 1.
In Experiment 2, we asked the participants to judge whether the stimulus face is
more or less similar to Bob or John. Although the two prototypes (Bob and John)
were never presented as stimuli, we named this process “identification”—as we
did in Experiment 1.

Results and Discussion

Mean probabilities of identifying “Bob.” In the children’s group the average
probabilities of identifying the faces as Bob were analyzed (see points in Fig. 4a).
Both half-face conditions, the lower part of the face (see left panel), F(2, 20) =
155.68, p < .01, and the upper part of the face (see right panel), F(2, 20) = 81.91,
p < .01, were effective in changing the identifications from Bob to John.
In the whole-face conditions (middle panel) the influence of the eyes was atten-
uated (mean probabilities: .63, .46, and .25) but, in contrast to Experiment 1, it
was influential, F(2, 20) = 14.6, p < .01. The lower part was also influential (mean
probabilities: .85, .32, and .18), F(2, 20) = 47, p < 01. There also was a statisti-
cally significant interaction between the variables, F(4, 40) = 2.92, p < .05. Thus,
by decreasing the salience of the mouth variations and the overall size of the stim-
ulus faces, children took into account the eyes information to identify the faces.
Note that the variations of the eyes (and their covariates) used in Experiment 2
were identical to those used in Experiment 1.
For the adult group the same pattern of the average probabilities of identifying
the faces as Bob appeared (see points in Fig. 4b): The face halves were influential
in the half-face (left and right panel) as well as in the whole-face conditions (mid-
dle panel). In the latter, the interaction was also significant (all Fs were significant
at the .01 level). In the whole-face conditions, however, the influence of the mouth
(mean probabilities: .57, .49, and .38), the more influential feature of the children,
was attenuated in comparison to the influence of the eyes (mean probabilities: .92,
.45, and .08). In comparing the average probabilities of the whole-face conditions
(depicted in the middle panels of Figs. 4a and 4b) for the children the steepness of
the curves (the influence of the mouth), whereas for the adults the spreadness of
the curves (the influence of the eyes) is pronounced. This result provides evidence
for developmental differences in the information used in face identification.
Model tests. Although both children and adults considered eyes and mouth in
identifying the whole faces, they weighted eyes and mouth differently. Therefore
we tested a biased or weighted Fuzzy Logical Model of Perception in addition to
the simple Fuzzy Logical Model of Perception and the Single Channel Model
against the results. In the weighted Fuzzy Logical Model of Perception, the con-
tribution of the less important information, i.e., the upper face for the children’s
group, is attenuated by some proportion in the whole-face condition relative to
the half-face condition. This model might be better able to describe the results
156 SCHWARZER AND MASSARO

than the simple Fuzzy Logical Model of Perception. The equation for the weight-
ed Fuzzy Logical Model of Perception is
fi(b) ⫽ wfi(h) ⫹ .5(1 ⫺ w),
where fi(b) is the feature value on trials of the whole-face conditions and wfi(h) is
the feature value on trials of the half-face condition. The w is a free parameter
indicating the relative amount of influence on trials of the whole-face conditions
relative to the half-face condition. Given the 3X3 expanded factorial design, in
the simple Fuzzy Logical Model of Perception six free parameters are necessary
to fit the Fuzzy Logical Model of Perception to the 15 conditions. In the weight-
ed Fuzzy Logical Model of Perception, given the additional weight parameter
seven free parameters are necessary.
The results of the children’s group confirmed our expectations that the weight-
ed Fuzzy Logical Model of Perception might be better able to describe the results
than the simple Fuzzy Logical Model of Perception as well as the Single Channel
Model. In comparison to the simple Fuzzy Logical Model of Perception and the
Single Channel Model, the weighted Fuzzy Logical Model of Perception yielded
the best fit of the data. This can be seen in the generally lower RMSDs of the
weighted Fuzzy Logical Model of Perception. The RMSDs of the weighted Fuzzy
Logical Model of Perception ranged between .01 to .10, with an average RMSD
of .06 and a fit of the mean participant data of .04. The RSMDs of the simple
Fuzzy Logical Model of Perception and the Single Channel Model were signifi-
cantly higher, F(2, 20) = 5.77, p < .01, (simple Fuzzy Logical Model of
Perception’s RMSD’s ranged from .02 to .13, with an average RMSD of .07 and
a fit of the mean participant data of .06, Single Channel Model’s RMSDs ranged
from .04 to .12, with an average RMSD of .08 and a fit of the mean participant of
.04). Thus, the good fit of the weighted Fuzzy Logical Model of Perception
(which can be seen by comparing the points and corresponding lines in Fig. 4a)
implicates the analytic as well as integrative character of the processing mode in
children’s face identification performance.
In contrast to the children, the adults were more influenced by the upper part
of the face in the whole-face conditions. We therefore contrasted in the adults’
group the corresponding weighted Fuzzy Logical Model of Perception with the
Single Channel Model and the simple Fuzzy Logical Model of Perception. As for
the children, the RMSDs of the weighted Fuzzy Logical Model of Perception
were significantly lower than those of the simple Fuzzy Logical Model of
Perception and the SCM, F(2, 24) = 5.19, p < .01, (weighted Fuzzy Logical
Model of Perception RMSDs ranged between .03 to .13 with an average RMSD
of .07 and a fit of the mean participant of .03, whereas the simple Fuzzy Logical
Model of Perception ranged from .03 to .15 with an average RMSD of .1 and a
fit of the mean participant of .07 and the Single Channel Model RMSDs ranged
from .04 to .13 with an average RMSD of .09 and a fit of the mean participant of
.06). The good fit of the weighted Fuzzy Logical Model of Perception in the
adults is demonstrated in Fig. 4b in that the predictions of the weighted Fuzzy
 FACE PROCESSING 157

Logical Model of Perception (see lines) did reasonably well in capturing the
trends in the data (see points).
In sum, the best fit of the weighted Fuzzy Logical Model of Perception impli-
cates that children’s face identification processing—like adults’—is based on the
independent evaluation of single facial features and can be named as analytic.
Taking the age-specific use of information into account, children’s processing
modes can also be described by a multiplicative integration of the facial features.
Although differences in the influence of the information sources were observed
in Experiments 1 and 2, there seem to be no differences in the face identification
processing between children and adults.

GENERAL DISCUSSION
The presented experiments proved successful in addressing the issues of how
two facial features are evaluated and integrated to achieve the identification of
a face. In adults as well as in children (see Experiment 2) both eyes and mouth
variations were effective in changing the judgment from the identification of
one face (Bob) to another face (John). These results were well described by the
Fuzzy Logical Model of Perception the weighted Fuzzy Logical Model of
Perception respectively relative to the poorer description of the Single Channel
Model. Thus, this is evidence for analytic face processing in children and
adults.
Given the present results that the Single Channel Model provided a poorer pre-
diction of our empirical data than the Fuzzy Logical Model of Perception, an
important question to consider is whether cases could be identified where the
Single Channel Model provides a better fit to the data than the Fuzzy Logical
Model of Perception. It has been claimed that when a face is turned upside down,
the integrative face processing strategy is inhibited, forcing a more analytic pro-
cessing strategy (e.g., Farah, Tanaka, Wilson, & Drain, 1995; Farah, Wilson,
Drain, & Tanaka, 1998; Tanaka & Farah, 1993). If this is true, the Single Channel
Model should provide a better account of the faces in an inverted condition than
the Fuzzy Logical Model of Perception. However, previous data already showed
that changing the information on a face—as you would do in presenting the faces
upside down—did not change the fit of the Fuzzy Logical Model of Perception
(e.g., Massaro, 1998; Massaro & Cohen, 1996). For example, Massaro and
Cohen (1996) examined whether the orientation of the face influences speech
perception. Inverting the face reduced the influence of visible speech, but the
results with both the upright and inverted faces were adequately described by the
Fuzzy Logical Model of Perception. The observed differences between the pro-
cessing of the upright and inverted faces were due to the amount of visible infor-
mation but inverting the faces did not change the nature of the information pro-
cessing. Thus, we might also expect that the amount of information on the faces
will be reduced in a face identification situation but with regard to the informa-
tion processing the Fuzzy Logical Model of Perception will do as well with
inverted faces.
158 SCHWARZER AND MASSARO

As already argued, the good fit of the Fuzzy Logical Model of Perception
challenges the conclusion of holistic face processing based on approaches with-
out any formal model testing. Whereas our results illustrated the independent
evaluation of facial features which stands for an analytic processing, Tanaka
and Farah (1993), Tanaka et al. (1998), Baenninger (1994), and Carey (1996)
postulated holistic face processing in children and adults in terms of holistic
encoding and configurational encoding. One possible explanation for the
results of the present study in comparison to the results of previous results
showing holistic processing in children and adults could lie in the different
number of target faces used in the different studies. Whereas in our study only
two target faces (Bob and John) were used previous face recognition studies
have used multiple target faces. Possibly, the use of only two target faces
encourage an analytic processing strategy because it is easier to focus on one
or two discriminating features when there are only two alternatives in the
response set. Holistic processing may be more likely to be employed in the sit-
uation of multiple target faces where the discriminating features are less appar-
ent. This conjecture is open to experimental test. In speech emotion perception,
the number of test and response alternatives does not diminish the advantage
of the description given by the Fuzzy Logical Model of Perception. Another
explanation for the difference between the present results and the previous ones
showing holistic face processing could be the fact that the target faces in the
previous studies were not present when the participant was deciding on the
identity of each stimulus face and he or she could only compare the target face
to a “blurry” memory representation. However, the stimulus faces were pre-
sented in our study for just 1500 ms so that a direct and nearly simultaneous
comparison of stimulus and target faces—which could possibly induce analyt-
ic processing—could not take place.
Our results confirm and extend current studies concerning children’s prefer-
ence to focus on single facial features when categorizing faces in comparison to
adults who took more than one feature into account (Pedelty, Levine, & Shevell,
1985; Schwarzer, 2000; Schwarzer & Korell, in press). However, with the appro-
priate balance in feature salience, both children and adults not only take more
than one feature into account but they integrate these features by using a multi-
plicative algorithm.
The present results are also consistent with the corresponding studies from the
nonfacial visual domain. Different tasks from the present one (Ashkenasy &
Odom, 1982; Thompson, 1994; Ward, Vela, & Hass, 1990; Wilkening & Lange,
1989) as well as comparable tasks (e.g., Thompson & Massaro, 1989) revealed the
analytic and integrative nature of processing. Given the success of the Fuzzy
Logical Model of Perception across a wide range of empirical domains, its suc-
cess in the present studies might not be too surprising. The Fuzzy Logical Model
of Perception has provided an adequate description of the evaluation and integra-
tion of sources of information in reading letters and words, in sentence process-
ing, in the visual perception of depth, in memory retrieval, and in cognitive deci-
 FACE PROCESSING 159

sion making (Massaro, 1998). It is also noteworthy that most of its success has
emerged in the description of speech perception (e.g., Massaro, 1987a; Massaro
& Burke, 1991). There has been a long tradition of belief that speech perception
is somehow specialized and not amenable to a description grounded in prototypi-
cal pattern recognition processes. This belief parallels the assumption of many that
the perception of faces is also specialized (see for an overview Bruce &
Humphreys, 1994). However, studies using the Fuzzy Logical Model of
Perception approach in the speech domain have weakened the speech-is-special
viewpoint (Massaro, 1987a, 1987b, 1998). We would predict the same for face
identification.
What do the current results tell us about the development of face identification
processing? Developmental changes could theoretically exist at any of the three
stages of processing: evaluation, integration, and decision. Given the fit of the
Fuzzy Logical Model of Perception to the children’s and adults’ data, however,
the results of the presented experiments provide evidence for developmental dif-
ferences only at evaluation. That means that differences in the evaluation process
appear to be due to differences in the information made available by the percep-
tual and memory system. Our results showed those differences between 5-year-
old children and adults in that in Experiment 1 and 2 children used the informa-
tion on the mouth of the faces more than adults did. Adults, by contrast were
more influenced by the eyes of the faces as can be seen in Experiment 2.
However, it would be oversimplified to conclude from our studies that in face
processing children usually prefer information on the mouth of the face and
adults usually prefer information on the eyes of a face. These differences seem
to depend not simply on the age but on differential salience of the features (see
Ashkenasy & Odom, 1982; Cook & Odom, 1992). The influence of the salience
of the features was observable in the way children changed their use of informa-
tion between Experiments 1 and 2. While, in terms of Cook and Odom (1992),
the predisposed salience of the mouth information (the effects that perceptual
experience with the mouth of a face has on the sensitivity of the perceptual sys-
tem) was higher than that of the eyes information and their covariates (eyebrows
and forehead) in Experiment 2, the salience of the mouth was reduced along with
the overall size of the faces, which increased the relative salience of the eyes and
their covariates and made them more informative.
Moreover, Experiment 2 showed that no developmental changes exist at the
stage of the integration of information from eyes and mouth. Rather, information
integration appears to occur in the same manner for children and adults: For both
age groups, the two sources of information are integrated by a multiplicative algo-
rithm.
Thus, in line with Cook and Odom’s (1992) differential-sensitivity account of
cognitive processing (see also Odom & Guzman, 1972), room for the develop-
ment of facial processing is mainly given in terms of facial information used
rather than in terms of information processing. As children become more expe-
rienced with the world around them, they learn what information on a face is
160 SCHWARZER AND MASSARO

useful. How they use this information to respond to their environment, however,
does not change.

REFERENCES
Aschkenasy, J. R., & Odom, R. D. (1982). Classification and perceptual development: Exploring
issues about integrality and differential sensitivity. Journal of Experimental Child Psychology,
34, 435–448.
Baenninger, M. A. (1994). The development of face recognition: Featural or configural processing.
Journal of Experimental Child Psychology, 57, 377–396.
Bruce, V., & Humphreys, G. W. (1994). Recognizing objects and faces. Visual Cognition, 1, 141–180.
Bushnell, I. W. (1982). Discrimination of faces by young infants. Journal of Experimental Child
Psychology, 33, 298–308.
Bushnell, I. W., Sai, F., & Mullin, J. T. (1989). Neonatal recognition of the mother’s face. British
Journal of Developmental Psychology, 7, 3–15.
Carey, S. (1996). Perceptual classification and expertise. In R. Gelman & T. Kit-Fong Au (Eds.),
Perceptual and cognitive development (pp. 49–69). San Diego, CA: Academic Press.
Carey, S., & Diamond, R. (1994). Are faces perceived as configurations more by adults than by chil-
dren? Visual Cognition, 1, 253–274.
Chandler, J. P. (1969). Subroutine STEPIT finds local minima of a smooth function of several param-
eters. Behavioral Science, 14, 81–82.
Cook, G. L., & Odom, R. D. (1992). Perception of multidimensional stimuli: A differential-sensitivi-
ty account of cognitive processing and development. Journal of Experimental Child Psychology,
54, 213–249.
Cunningham, J. G., & Odom, R. D. (1978). The role of perceptual salience in the development of
analysis and synthesis process. Child Development, 49, 815–823.
Cunningham, J. G., & Odom, R. D. (1986). Differential salience of facial features in children’s per-
ception of affective expression. Child Development, 57, 136–142.
Diamond, R., & Carey, S. (1986). Why faces are and are not special: An effect of expertise. Journal
of Experimental Psychology: General, 115, 107–117.
Farah, M. J., Tanaka, J. W., & Drain, H. M. (1995). What causes the face inversion effect? Journal of
Experimental Psychology: Human Perception and Performance, 21, 628–634.
Farah, M. J., Wilson, K. D., Drain, M., & Tanaka, J. N. (1998). What is special about face perception.
Psychological Review, 105, 482–498.
Flin, R., & Dziurawiec, S. (1989). Developmental factors in face processing. In A. W. Young & H.
D. Ellis (Eds.), Handbook of research on face processing (pp. 335–378). Amsterdam: Elsevier.
Garner, W. R. (1974). The processing of information and structure. Hillsdale, NJ: Erlbaum.
Kemler Nelson, D. G. (1989). The nature and occurrence of holistic processing. In B. E. Shepp & S.
Ballesteros (Eds.), Object perception: Structure and process (pp. 357–386). Hillsdale, NJ:
Erlbaum.
Massaro, D. W. (1987a). Speech perception by eye and ear: A paradigm for psychological inquiry.
Hillsdale, NJ: Erlbaum.
Massaro, D. W. (1987b). Categorical partition: A fuzzy logical model of categorization behavior. In
S. Harnad (Ed.), Categorical perception: The groundwork of cognition (pp. 254–283). New York:
Cambridge Univ. Press.
Massaro, D. W. (1998). Perceiving talking faces. Cambridge, MA: MIT Press.
Massaro, D. W., & Burke, D. (1991). Perceptual development and auditory backward recognition
masking. Developmental Psychology, 27, 85–96.
Massaro, D. W., & Cohen, M. M. (1996). Perceiving speech from inverted faces. Perception &
Psychophysics, 58, 1047–1065.
Massaro, D. W., & Friedman, D. (1990). Models of integration given multiple sources of information.
Psychological Review, 97, 225–252.
 FACE PROCESSING 161

Odom, R. D., & Guzman, R. D. (1972). Development of hierarchies of dimensional salience.

Developmental Psychology, 6, 271–287.
Pascalis, O., de Haan, M., Nelson, C. A., & de Schoenen, S. (1998). Long-term recognition memory
for faces assessed by visual paired comparison in 3- and 6-month- old infants. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 24, 249–260.
Pedelty, L., Levine, S., & Shevell, S. K. (1985). Developmental changes in face processing: Results
from multidimensional scaling. Journal of Experimental Child Psychology, 39, 421–436.
Schwarzer, G. (1997). Analytic and holistic modes in the development of melody perception.
Psychology of Music, 25, 35–56.
Schwarzer, G. (2000). Development of face processing: The effect of face inversion. Child
Development, 71, 391–401.
Schwarzer, G., & Korell, M. (in press). Gesichterverarbeitung im Kleinkind- und Vorschullalter.
Zeitschrift für Entwicklungspsychologie und Pädagogische Psychologie.
Shepp, B. E. (1978). From perceived similarity to dimensional structure. In E. Rosch & B. B. Lloyd
(Eds.), Cognition and categorization (pp. 135–167). Hillsdale, NJ: Erlbaum.
Smith, L. B., & Kemler, D. G. (1977). Developmental trends in free classification: Evidence for a new
conceptualization of perceptual development. Journal of Experimental Child Psychology, 24,
279–298.
Tanaka, J. W., & Farah, M. J. (1993). Parts and wholes in face recognition. Quarterly Journal of
Experimental Psychology, 46A, 225–245.
Tanaka, J. W., Kay, J. B., Grinell, E., Stansfield, B., & Szechter, L. (1998). Face recognition in young
children: When the whole is greater than the sum of its parts. Visual Cognition, 5, 479–496.
Tanaka, J. W., & Sengco, J. A. (1997). Features and their configuration in face recognition. Memory
& Cognition, 25, 583–592.
Thompson, L. A. (1994). Dimensional strategies dominate perceptual categorization. Child
Development, 65, 1627–1645.
Thompson, L. A., & Massaro, D. W. (1989). Before you see it, you see its parts: Evidence for feature
encoding and integration in preschool children and adults. Cognitive Psychology, 21, 334–362.
Vetter, T., & Troje, N. F. (1997). Separation of texture and shape in images of faces for image encod-
ing and synthesis. Journal of the Optical Society of America, 14, 2152–2161.
Ward, T. B., Vela, E., & Hass, S. D. (1990). Children and adults learn family-resemblance categories
analytically. Child Development, 61, 593–605.
Werner, H. (1926). Einführung in die Entwicklungspsychologie. Leipzig: Barth.
Wilkening, F., & Lange, K. (1989). When is children’s perception holistic? Goals and styles in pro-
cessing multidimensional stimuli. In T. Globerson & T. Zelnicker (Eds.), Cognitive style and
cognitive development (pp. 141–171). Norwood, NJ: Ablex.
Young, A. W., Hellawell, D., & Hay, D. C. (1987). Configural information in face perception.
Perception, 16, 747–759.

Received October 1, 1998; revised April 4, 2000